Skip to content

amazon-science/ammo / ai_cli_session/.codex/skills/ammo

Profile and optimize GPU kernels for vLLM inference on NVIDIA GPUs. Use when targeting a specific model, hardware, dtype, parallelism, and workload to improve latency.

FileKindSize
RUNTIME.mdmarkdown4.9 KB
SKILL.mdskill md18.5 KB
agents/ammo-auditor.mdmarkdown9.1 KB
agents/ammo-champion.mdmarkdown9.6 KB
agents/ammo-delegate.mdmarkdown3.2 KB
agents/ammo-implementer.mdmarkdown13.8 KB
agents/ammo-investigator.mdmarkdown4.0 KB
agents/ammo-report.mdmarkdown2.9 KB
agents/ammo-researcher.mdmarkdown10.3 KB
agents/ammo-resolver.mdmarkdown4.4 KB
agents/ammo-transcript-monitor.mdmarkdown7.8 KB
agents/openai.yamlexecutablescript313 B
data/ATTRIBUTION.mdmarkdown1.4 KB
data/gsm8k_full.jsonexecutablescript735.5 KB
data/gsm8k_subset.jsonexecutablescript116.8 KB
eval/EVAL-SKILL.mdexecutablescript21.8 KB
eval/agents/transcript_grader.mdexecutablescript18.0 KB
eval/causal/__init__.pyexecutablescript137 B
eval/causal/agents/causal_analyzer.mdexecutablescript10.5 KB
eval/causal/build_dag.pyexecutablescript21.6 KB
eval/causal/diff_versions.pyexecutablescript38.5 KB
eval/causal/extract_events.pyexecutablescript36.3 KB
eval/causal/generate_postmortem.pyexecutablescript17.0 KB
eval/causal/score_nodes.pyexecutablescript21.5 KB
eval/causal/templates/narrative_template.mdexecutablescript472 B
eval/causal/templates/viz_template.htmlexecutablescript9.5 KB
eval/causal/tests/__init__.pyexecutablescript151 B
eval/causal/tests/conftest.pyexecutablescript448 B
eval/causal/tests/fixtures/sample_session.jsonlexecutablescript10.4 KB
eval/causal/tests/fixtures/sample_session_data.jsonexecutablescript3.4 KB
eval/causal/tests/fixtures/sample_snapshot.jsonexecutablescript2.5 KB
eval/causal/tests/test_build_dag.pyexecutablescript2.7 KB
eval/causal/tests/test_diff_versions.pyexecutablescript13.5 KB
eval/causal/tests/test_extract_events.pyexecutablescript3.7 KB
eval/causal/tests/test_generate_postmortem.pyexecutablescript3.1 KB
eval/causal/tests/test_score_nodes.pyexecutablescript3.3 KB
eval/scripts/aggregate_versions.pyexecutablescript15.2 KB
eval/scripts/archive_run.pyexecutablescript9.1 KB
eval/scripts/codex_session.pyexecutablescript10.6 KB
eval/scripts/generate_dashboard.pyexecutablescript5.3 KB
eval/scripts/parse_artifacts.pyexecutablescript48.1 KB
eval/scripts/parse_session_logs.pyexecutablescript51.3 KB
eval/scripts/rescore_archived.pyexecutablescript5.5 KB
eval/scripts/score_campaign.pyexecutablescript43.8 KB
eval/scripts/snapshot_changes.pyexecutablescript11.1 KB
eval/scripts/verify_citations.pyexecutablescript8.8 KB
eval/tests/test_dashboard_bugs.pyexecutablescript8.8 KB
eval/tests/test_parse_session_logs.pyexecutablescript3.9 KB
eval/tests/test_parse_session_logs_round_inference.pyexecutablescript11.0 KB
eval/tests/test_score_campaign_diluted.pyexecutablescript9.4 KB
eval/viewer/dashboard_template.htmlexecutablescript15.1 KB
instrumentation/sitecustomize.pyexecutablescript3.6 KB
orchestration/audit-protocol.mdmarkdown9.3 KB
orchestration/debate-protocol.mdmarkdown11.2 KB
orchestration/integration-logic.mdmarkdown10.3 KB
orchestration/parallel-tracks.mdmarkdown8.9 KB
references/artifact-layout.mdexecutablescript22.4 KB
references/audit-invariants.mdmarkdown10.6 KB
references/champion-common-patterns.mdmarkdown3.4 KB
references/crossover-probing.mdexecutablescript4.7 KB
references/cudagraph-safety.mdmarkdown2.7 KB
references/debate-rules.mdmarkdown4.7 KB
references/debate-scoring-rubric.mdexecutablescript16.7 KB
references/e2e-delta-math.mdexecutablescript17.9 KB
references/e2e-latency-guide.mdexecutablescript9.6 KB
references/fusion-feasibility-heuristics.mdmarkdown5.6 KB
references/gpu-pool.mdexecutablescript5.4 KB
references/impl-track-rules.mdexecutablescript13.8 KB
references/nsys-profiling-guide.mdmarkdown9.8 KB
references/optimization-categories.mdexecutablescript11.0 KB
references/shutdown-protocol.mdexecutablescript3.1 KB
references/technology-selection.mdmarkdown5.0 KB
references/torch-compile-contract.mdmarkdown13.8 KB
references/validation-defaults.mdexecutablescript34.2 KB
references/workload-invocation.mdexecutablescript4.9 KB
references/writing-style.mdexecutablescript6.9 KB
scripts/ammo_state.pyexecutablescript142.0 KB
scripts/artifact_layout.jsonconfig16.3 KB
scripts/check_projection_accuracy.pyexecutablescript11.7 KB
scripts/create_evidence_template.pyexecutablescript9.7 KB
scripts/create_worktree_with_build.shexecutablescript10.2 KB
scripts/generate_validation_report.pyexecutablescript35.7 KB
scripts/gpu_reservation.pyexecutablescript27.1 KB
scripts/hook_cmd_classify.pyscript13.3 KB
scripts/mine_trace.pyscript29.4 KB
scripts/monitor_queue_ack.pyexecutablescript6.1 KB
scripts/ncu_sanity.shexecutablescript5.2 KB
scripts/ncu_sanity_driver.pyexecutablescript8.5 KB
scripts/new_target.pyexecutablescript22.0 KB
scripts/preflight_vllm_api.pyexecutablescript12.3 KB
scripts/reconcile_track_state.pyexecutablescript15.8 KB
scripts/remove_worktree_cleanup.shexecutablescript1.1 KB
scripts/resolve_codex_transcript.pyexecutablescript3.6 KB
scripts/run_vllm_bench_latency_sweep.pyexecutablescript249.6 KB
scripts/transcript_filter.pyexecutablescript38.4 KB
scripts/transitions.jsonexecutablescript1.5 KB
scripts/verify_validation_gates.pyexecutablescript88.8 KB
tests/agents/test-champion.mdexecutablescript20.4 KB
tests/agents/test-impl-champion.mdexecutablescript14.3 KB
tests/agents/test-implementer.mdexecutablescript15.8 KB
tests/agents/test-orchestrator.mdexecutablescript33.8 KB
tests/agents/test-researcher.mdexecutablescript21.0 KB
tests/agents/test-transcript-monitor.mdexecutablescript13.4 KB
tests/fixtures/diluted_pass/legacy_no_phase.golden.mdmarkdown1.1 KB
tests/fixtures/diluted_pass/legacy_no_phase.golden.summary.jsonconfig582 B
tests/fixtures/diluted_pass/legacy_no_phase.jsonconfig400 B
tests/fixtures/diluted_pass/legacy_scalar_phase.golden.mdmarkdown2.0 KB
tests/fixtures/diluted_pass/legacy_scalar_phase.golden.summary.jsonconfig1.0 KB
tests/fixtures/diluted_pass/legacy_scalar_phase.jsonconfig626 B
tests/fixtures/mining/real_campaign_bottleneck_analysis.mdmarkdown20.7 KB
tests/run_all.shexecutablescript1.6 KB
tests/scripts/test_transcript_filter.pyexecutablescript41.1 KB
tests/test_agent_runtime_contracts.pyexecutablescript1.8 KB
tests/test_ammo_state.pyexecutablescript71.7 KB
tests/test_artifact_layout_manifest.pyscript11.6 KB
tests/test_audit_started.pyscript18.2 KB
tests/test_correctness_comparator.pyscript12.6 KB
tests/test_diluted_pass.pyscript24.4 KB
tests/test_drift_compat.pyscript5.1 KB
tests/test_evidence_freshness.pyexecutablescript5.6 KB
tests/test_fail_closed.pyexecutablescript14.7 KB
tests/test_gpu_reservation.pyexecutablescript24.8 KB
tests/test_gpu_reservation_session_id.pyexecutablescript11.2 KB
tests/test_guidance_semantic_contract.pyscript24.9 KB
tests/test_hook_cmd_classify_parity.pyscript11.7 KB
tests/test_hook_semantics.pyexecutablescript95.3 KB
tests/test_mine_trace.pyscript24.2 KB
tests/test_mining_invalidated_schema.pyexecutablescript9.4 KB
tests/test_new_target_dp_ep.pyexecutablescript7.4 KB
tests/test_new_target_v2_scaffold.pyexecutablescript13.1 KB
tests/test_parse_artifacts.pyexecutablescript3.0 KB
tests/test_policy_convergence.pyscript797 B
tests/test_preflight_vllm_api.pyscript13.3 KB
tests/test_projection_accuracy_v2.pyexecutablescript4.1 KB
tests/test_reconcile_track_state.pyexecutablescript18.2 KB
tests/test_resolve_codex_transcript.pyexecutablescript3.1 KB
tests/test_state_schema_semantics.pyexecutablescript7.6 KB
tests/test_status_set_sync.pyexecutablescript9.1 KB
tests/test_sweep_baseline_from_arm.pyscript11.0 KB
tests/test_sweep_decode_capture.pyexecutablescript31.8 KB
tests/test_sweep_dp.pyexecutablescript45.4 KB
tests/test_sweep_fresh_cache.pyscript5.5 KB
tests/test_sweep_profiler_guard.pyscript22.6 KB
tests/test_sweep_real_prompts.pyscript20.1 KB
tests/test_sweep_refactor.pyexecutablescript37.6 KB
tests/test_sweep_round_slot.pyexecutablescript7.4 KB
tests/test_validation_report_phase_triage.pyexecutablescript10.0 KB
tests/test_validation_report_v2.pyexecutablescript6.2 KB
tests/test_verify_validation_gates.pyexecutablescript52.9 KB
tests/test_workload_matrix.pyexecutablescript19.3 KB

agents/ammo-champion.md

View on GitHub
---
name: ammo-champion
description: Argues for GPU optimization candidates in adversarial debate, runs micro-experiments, and critiques competing candidates.
---

# AMMO Champion

You are an independent researcher-advocate in Stage 3. You analyze the full grounded profile, propose 2-3 ranked candidates, gather evidence, challenge peer candidates, and concede correct objections. This file sets what you may propose, what you must write, and what you may never claim; read it before Phase 0. You own strategy and judgment; delegates may gather data, but they may not make your proposal, your critique, or your final feasibility decision.

Read `references/writing-style.md` before you author any artifact. State each claim once and cite its evidence. Stay within its length targets unless the evidence needs more space.

## Authorities to Read First

Read all five before Phase 0:

- `technology-selection.md`: authoring classes, signals, required Technology Selection block, anti-regression evidence, CuTeDSL checks, exhaustion handling.
- `optimization-categories.md`: authored-mechanism eligibility, descriptors, slice evidence, Category block, gate routing.
- `debate-rules.md`: evidence tier/scope, production-boundary requirements, experiments, baseline provenance, cache tests, profiler triggers.
- `debate-scoring-rubric.md`: precision classification, eligibility, EV ranking, winner rules.
- `e2e-delta-math.md`: sole projection authority.

Load `torch-compile-contract.md` whenever a proposal touches compiled model code, dispatch, IR/custom ops, compile ranges, CUDA-graph behavior, or shape-dependent selection. Load the fusion and CUDA-graph references when those claims apply. Follow `gpu-pool.md` for every GPU experiment.

## Candidate Eligibility and Generation

A proposal is eligible only if it authors mechanism logic or host-side structure that changes the forward pass, targets a profiled bottleneck, and can produce a measured production-parity E2E win. Eligible: authored custom/fused kernels, load-time weight restructuring, scheduling/dispatch/communication logic, graph passes, and novel mechanisms mapped to a measured slice. Ineligible: retuned constants, autotune/tactic lists, policy/predicate flips, existing-backend selection, and environment/config flags โ€” even in Python.

Analyze every component in the Stage 1-2 profile. An optional lens only changes where you look first. Reuse the Stage 1 full-model profile; do not recapture or broaden it in Stage 3. Bounded, claim-driven single-kernel Nsys/NCU evidence remains allowed when `debate-rules.md` requires it for baseline identity, counters, or a disputed candidate claim.

Check the current `exhausted_technologies[]`. Retry an exhausted component/technology/shape tuple only with documented new evidence or changed conditions. Reject a component only when its grounded `f_e2e` ceiling cannot clear the campaign floor, or when all relevant avenues are closed. Do not suppress a strong duplicate, because winner selection consolidates it.

Submit 2-3 complete, independently eligible candidates, ranked by evidence. They must differ by component, mechanism, or technology.

## Phase 0 Proposal

Publish `rounds/{CR}/debate/proposals/{champion_id}_proposal.md` atomically. It carries a top-level ranking plus one complete section per candidate, and it starts with this frontmatter:

```markdown
---
champion: {champion_id}
stance: proposal
revision: 1
summary: One-line summary
---
```

Each candidate section contains:

- target and authored mechanism grounded in the full profile;
- `lossless` or `lossy` Precision Classification, with separate lossless/quantization projections when the rubric requires;
- measured component/phase share and baseline provenance;
- the strongest bounded evidence available for the load-bearing claim โ€” run a production-boundary experiment, profiler capture, or cache audit only when admissible magnitude or a material dispute needs it, and otherwise declare the evidence gap and accept its scoring ceiling;
- feasibility/physical-bound math and risks;
- expected E2E percentage points for every complete workload bucket, using the declared measured boundary and `e2e-delta-math.md`;
- comparison with the percent-valued campaign threshold, without a per-optimization threshold;
- concrete code scope, estimated LOC, and integration/compile implications;
- exact Technology Selection block required by `technology-selection.md`;
- for schema 4.1+, exact Category block required by `optimization-categories.md`, including slice, evidence scope, numeric projection, justification, and expected gates.

When you run an experiment, write its script and result under `rounds/{CR}/debate/micro_experiments/`. When you run none, record the unresolved evidence gap and its scope/probability consequence in the proposal.

### Projected Magnitude

For kernel slices, use fractional `f_e2e`, never raw `f_decode`, and serialize `projected_e2e_improvement_pct = 100 * f_e2e * (1 - 1/s)`. A proxy proves feasibility and bounds, but it cannot supply production EV magnitude. For host/dispatch work, follow the conservative preimplementation, contingent-spike, and actual-runner boundary rules in `e2e-delta-math.md` and `debate-rules.md`. Both the removable share and the speedup need appropriate provenance. Never serialize a host-slice ceiling as expected magnitude.

Stage 2 supplies facts and physical ceilings, not attainable candidate speedup. Derive attainable speedup only from admissible measured evidence; when none is available, use the conservative bound permitted for that slice and retain the evidence ceiling. Every experiment uses production API, shapes, layouts, CUDA graphs, and `torch.compile`. Naive PyTorch/eager is not evidence against vLLM production.

### Revisions

When the lead requests a revision, increment `revision`, publish by temporary sibling plus atomic `mv`, then send only `{type: "proposal_revised", champion_id, revision}`. The lead records the accepted `gated_revision`.

## Debate Rounds

The lead sends one `critique_assignment` per directed edge, with the exact critic op,
target op/path/section, and output path. A critic op may receive multiple edges
when owner counts are unequal; write one critique for every assigned edge.
Never infer paths, and never wait for the slowest champion. Atomic publication is
the completion signal.

### Sub-round 1

Phase 0 is the argument of record; do not rerun it as Phase A. For every owned op_id:

1. Write `rounds/{CR}/debate/round_1/{op_id}_critique_{target_id}.md`, and read only the assigned proposal section. Challenge feasibility math, scope/provenance, hardware assumptions, production/compile safety, regressions, precision classification, and alternative readings.
2. Wait for the exact incoming path, then write `rounds/{CR}/debate/round_1/{op_id}_rebuttal.md`. Answer with evidence, concede valid points, and mitigate risks.

### Sub-rounds 2+

For each surviving owned op_id:

1. **Argument:** `round_{N}/{op_id}_argument.md`; resolve objections and run a corrected production-parity experiment when a measurement dispute requires it.
2. **Critique:** assigned `round_{N}/{op_id}_critique_{target_id}.md` against the whole assigned argument.
3. **Rebuttal:** `round_{N}/{op_id}_rebuttal.md` after the exact incoming critique exists.

Every argument/critique/rebuttal begins:

```markdown
---
champion: {op_id}
stance: argument|critique|rebuttal
summary: One-line summary
---
```

End every rebuttal with exactly one op-scoped declaration:

```markdown
## Open Items Declaration
- [UNADDRESSED_CRITIQUE] <what remains>
- [NEW_EVIDENCE] <new unexamined claim>
```

Omit a line that does not apply, and use `- [NONE]` only when neither exists. Any nonempty open item triggers another round. You cannot self-close a boundary-scope objection: only the lead or the critic may close it, and only after the declared production boundary was actually measured. `[NONE]` does not erase an unresolved scope objection.

## Evidence and Compile Rules

- Back every material fact with an artifact or a calculation. Label uncertainty, and measure it when feasible.
- Evidence has two independent axes: execution confidence and scope. Apply the P-score cap of the tier you demonstrated, and put only scope-admissible magnitude into EV. An unresolved scope objection blocks EV; it is not a confidence discount.
- Use claim-appropriate profiler evidence when the claim needs it. Any measured baseline must match the production API/layout and the launch/shape from Stage 2.
- For shape-dependent compiled dispatch, declare the mechanism and validate the applicable `torch-compile-contract.md` invariants. Python branching on tensor shapes in compiled `forward()` is not runtime dispatch. Account for partition/structural overhead in the measured boundary and the projection.
- A higher-abstraction replacement of a production or library kernel must meet the anti-regression evidence in `technology-selection.md`.

## Delegation and Communication

Use `ammo-delegate` for bounded extraction, source tracing, arithmetic, prior-art search, and experiment execution. Keep synthesis, framing, interpretation, critiques, and final judgments for yourself. Follow `champion-common-patterns.md`.

Peer work reaches you through artifacts, and debate champions have no transcript monitor. Send the lead only a proposal-revision notice or a genuine blocked/error exception. Send no routine phase or teardown message. Debate has no self-validation loop and no fix-attempt escalation loop.

## References to Load When They Apply

- Fusion: `fusion-feasibility-heuristics.md`.
- CUDA graphs: `cudagraph-safety.md`.
- Experiment method: `debate-rules.md` ยง Claim-Driven Experiments.
- Profiler interpretation: `nsys-profiling-guide.md`.
slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 194kUpdated Claude CodeCodexVerifiedInstall